Staged Training Final Summary

Total Training Time

Duration: 18:18:34

Best Checkpoint

Name: best_model_auto_session_so101_should_pan_500_stage10_train_304_run2_longer_val_plateau_07484400_cont_val_0.035202.pth
Stage: 10
Hybrid Loss (full session): 0.036726

Learning Rate Schedule (Best Stage)

Stage Progression

Stage Orig Loss Train Loss Time Samples Stop Reason
1 0.105523 0.104704 00:00:41 4536 divergence
2 0.094195 0.078369 00:10:28 94752 divergence
3 0.067060 0.035586 05:54:15 3170664 divergence
4 0.065550 0.042827 00:00:34 4536 divergence
5 0.065729 0.098630 00:00:34 4536 divergence
6 0.056086 0.059230 00:01:17 10584 divergence
7 0.048483 0.047198 01:29:05 779688 divergence
8 0.040387 0.038710 05:01:35 2365776 divergence
9 0.038272 0.037149 01:53:20 859824 val_plateau
10 ⭐ 0.036726 0.035202 02:15:02 1076544 val_plateau

Hybrid Loss Over Original Session (per Stage)

Stage 1 - Hybrid Loss: 0.105523

Stage 2 - Hybrid Loss: 0.094195

Stage 3 - Hybrid Loss: 0.067060

Stage 4 - Hybrid Loss: 0.065550

Stage 5 - Hybrid Loss: 0.065729

Stage 6 - Hybrid Loss: 0.056086

Stage 7 - Hybrid Loss: 0.048483

Stage 8 - Hybrid Loss: 0.040387

Stage 9 - Hybrid Loss: 0.038272

Stage 10 (Best) - Hybrid Loss: 0.036726

Staged vs Baseline Comparison

Overall Winner: Baseline Training
Stages Won (Staged): 3
Stages Won (Baseline): 7
Average Improvement Ratio: 0.970 (>1 = staged better)
Average Improvement: -0.000399 (positive = staged better)

Progression Comparison

Per-Stage Comparison

Stage Staged Loss Baseline Loss Improvement Ratio Staged Samples Baseline Samples Winner
1 0.105523 0.104848 -0.000674 0.994 4,536 4,536 Baseline
2 0.094195 0.075850 -0.018345 0.805 94,752 76,608 Baseline
3 0.067060 0.074402 +0.007343 1.109 3,170,664 84,672 Staged
4 0.065550 0.071843 +0.006294 1.096 4,536 70,056 Staged
5 0.065729 0.102942 +0.037214 1.566 4,536 41,328 Staged
6 0.056086 0.052040 -0.004046 0.928 10,584 82,656 Baseline
7 0.048483 0.042098 -0.006386 0.868 779,688 92,736 Baseline
8 0.040387 0.032519 -0.007868 0.805 2,365,776 110,376 Baseline
9 0.038272 0.028999 -0.009273 0.758 859,824 105,840 Baseline
10 0.036726 0.028474 -0.008252 0.775 1,076,544 93,744 Baseline

Sample Counts

Cumulative Across All Stages

Per Stage

Stage 1 - Total Samples: 4,536

Stage 2 - Total Samples: 94,752

Stage 3 - Total Samples: 3,170,664

Stage 4 - Total Samples: 4,536

Stage 5 - Total Samples: 4,536

Stage 6 - Total Samples: 10,584

Stage 7 - Total Samples: 779,688

Stage 8 - Total Samples: 2,365,776

Stage 9 - Total Samples: 859,824

Stage 10 (Best) - Total Samples: 1,076,544

Best Checkpoint Inference

Selected Frame 3

Action 0

Action 1

Action 2

Random Observations

Observation 274

Action 0
Action 1
Action 2

Observation 356

Action 0
Action 1
Action 2